We have identified an issue during our edge case testing where the model overpredicts risk when you have a flat spatial distribution and more tightly clustered turbines.
(EDW == effective detection (half) width and EDR == Effective detection radius)
In the point count example - we calculate the EDR and the observed flux window and then scale this up to a turbine level. This assumes that the observed flux can be applied to each turbine individually.
However, if the turbine layout is more compact and/or your EDR is large then the effective search area contains more then one turbine, so the observed flux needs to be weighted for the effective turbines in the search area. You can do this by calculating the total turbine observation area - say union_area_t = spatial union of turbines with a buffer of radius EDR) and the sum of individual buffers (sum_area_t = sum[ (pi*EDR^2) * n_turbines ])
The cluster correction factor is sum_area_t / union_area_t which is always >= 1
For line transects or first order approx to the point transect example you can do a linear correction of 2*ESW / mean(min_dist) where min_dist is the distance of each turbine to its nearest neighbour
This slots into the process by correcting the observed_flux when scaling up to a turbine. If you calculate the actual spatial 2D probability density using kernel density estimation or some other method, you would use that correction at the same place in the process to adjust the observed_flux by the probability a flight will occur through that turbine. The cluster correction is just the correct spatial probablity in the case of an assumption of a 'flat' (equal probability) flux across the site
To do this I'm going to make a linear and area correction function as above and maybe a control function where you can choose which method.
this issue is just to get the functionality in and tested (and needs to have #40 and #41 done first preferably). Then I'll make a second issue to rejig the vignettes so its all really clear where these new changes go.
This has been tested against our simulation and internally on test data based on sites with well understood collision risk.
We have identified an issue during our edge case testing where the model overpredicts risk when you have a flat spatial distribution and more tightly clustered turbines.
(
EDW== effective detection (half) width andEDR== Effective detection radius)In the point count example - we calculate the
EDRand the observed flux window and then scale this up to a turbine level. This assumes that the observed flux can be applied to each turbine individually.However, if the turbine layout is more compact and/or your
EDRis large then the effective search area contains more then one turbine, so the observed flux needs to be weighted for the effective turbines in the search area. You can do this by calculating the total turbine observation area - sayunion_area_t= spatial union of turbines with a buffer of radiusEDR) and the sum of individual buffers (sum_area_t = sum[ (pi*EDR^2) * n_turbines ])The cluster correction factor is
sum_area_t / union_area_twhich is always >= 1For line transects or first order approx to the point transect example you can do a linear correction of
2*ESW / mean(min_dist)wheremin_distis the distance of each turbine to its nearest neighbourThis slots into the process by correcting the
observed_fluxwhen scaling up to a turbine. If you calculate the actual spatial 2D probability density using kernel density estimation or some other method, you would use that correction at the same place in the process to adjust theobserved_fluxby the probability a flight will occur through that turbine. The cluster correction is just the correct spatial probablity in the case of an assumption of a 'flat' (equal probability) flux across the siteTo do this I'm going to make a linear and area correction function as above and maybe a control function where you can choose which method.
this issue is just to get the functionality in and tested (and needs to have #40 and #41 done first preferably). Then I'll make a second issue to rejig the vignettes so its all really clear where these new changes go.
This has been tested against our simulation and internally on test data based on sites with well understood collision risk.